Program understanding aims at discovering human-readable properties of a software project from the analysis of its source code. Recently, we proposed a smart approach based on hierarchical agglomerative clustering that extracts so-called program topoi from source code. These topoi are high-level observable properties of the project. Based on textual and structural representations of the source code, our multi-steps approach clusters program topoi in an effective and efficient way. In this paper, we depict novel exploitation tasks of this program understanding approach and report on its application to Software Heritage. Software Heritage is an ambitious project which aims at collecting and archiving the biggest corpus of publicly available software source code. One of the project goals is to provide a new scientific instrument for computer scientists to evaluate advanced machine learning and software engineering methods on a very large source code repository. Our in-depth experiments reveal that unsupervised learning is the appropriate tool to mine and understand the biggest corpus of software source code ever produced.